{"slug": "show-hn-hordev-claude-code-skills-that-build-instead-of-asking-questions", "title": "Show HN: Hordev – Claude Code skills that build instead of asking questions", "summary": "Hordev, a skills library for Claude Code, enables parallel, fast prototype building by skipping approval gates and using cheap agents for volume work and strong models for judgment. It is a descendant of superpowers and targets exploratory development where multiple approaches need testing, while superpowers remains better for structured, spec-driven work.", "body_md": "A skills library for [Claude Code](https://claude.com/claude-code) that trades\ninterviewing for building.\n\nhordev takes a small amount of input, decides the rest, writes a spec and a test-driven design without asking anyone to approve them, and then throws a horde of cheap fast agents at the problem in parallel. The goal is a running prototype you can judge, quickly, across several directions at once.\n\nIt is a descendant of [superpowers](https://github.com/obra/superpowers) and\nkeeps its tenets: skills as version-controlled markdown, tests before\nimplementation, artifacts that outlive the session. It diverges on pacing.\n\nThey are aimed at different jobs, and hordev is not a strict upgrade.\n\n**Use superpowers** when structure is what you need: spec-driven development,\na framework to work inside, long autonomous runs on well-understood work, and\nreview checkpoints you actually want. That structure is where superpowers earns\nits keep, and hordev deliberately gives some of it up.\n\n**Use hordev** when you need to find out fast, and broadly: you are not sure\nwhat to build yet, several approaches look plausible, and the cheapest way to\nlearn is to have working versions of them in front of you. Breadth is the point\nas much as speed — a horde is good at exploring four directions at once, not\njust at doing one direction faster.\n\nIf you are executing a spec you already trust, hordev's speed buys you little and costs you the checkpoints. Use superpowers.\n\nEvery hordev run moves through five stages:\n\n| Stage | Skill | Model | \n|---|---|---|\n| Extract | `rapid-spec` | haiku | \n| Design | `writing-tdds` | haiku | \n| Cut | `decomposing-for-hordes` | opus | \n| Swarm | `dispatching-hordes` ,`reconciling-horde-output` | haiku fan-out, opus merge | \n| Verify | `horde-qa` | opus | \n\nWhen the right approach is genuinely unknown, the horde goes wide instead of\ndeep: `racing-prototypes` builds two to four competing versions at once, under\nidentical budgets and pre-committed judging criteria, and picks a winner on\nevidence from running code.\n\nSupporting skills: `prototype-first`, `racing-prototypes`, `assumption-ledger`,\n`debugging-in-a-horde`, `isolating-horde-workspaces`, `improving-hordev`,\n`writing-hordev-skills`. `using-hordev` is the entrypoint, injected at session\nstart by a hook.\n\n**No approval gate.** hordev never asks you to sign off on a spec or a TDD. It\ndecides, writes them, and starts building. This is the whole speed bet, and it\nis why QA is the most important skill in the library — with nobody approving\nthe design, verification is the only thing between a wrong spec and a wrong\nprototype.\n\n**Every question you did not get asked is written down.** Not asking is only\nhonest if the decision is visible, so each one becomes an entry in\n`.hordev/assumptions.md` with its blast radius and how it will be falsified.\nYou see the open ones at the end of a run. That ledger is the contract that\nmakes skipping the interview defensible instead of reckless.\n\n**Cheap agents do the volume, strong agents do the judgment.** Specs, TDDs, and\nimplementation run on Haiku; orchestration, decomposition, reconciliation, and\nQA run on Opus. Reaching for the strongest model everywhere defeats the purpose\nof a horde.\n\nRuns append to `.hordev/run-log.md`. A `Stop` hook notices when the log has\ngrown and asks whether anything should change a skill. The bar is deliberately\nhigh: amend on a repeated failure, never on a one-off. `improving-hordev` maps\nsymptoms back to the skill that probably needs the edit — seam bugs usually\nmean the decomposition rules are too loose, agents returning plans instead of\ncode mean the dispatch prompt contract is weak.\n\nThe skills drop the occasional orcish line — \"Zug zug\" taking the job, \"For the Horde!\" when a wave dispatches, \"Work complete!\" when the tree comes back together. It is a running joke with rules: conversational output only, never two messages in a row, and never on top of bad news. Specs, TDDs, commit messages, and QA findings stay clean, because those outlive the joke.\n\nIn Claude Code, add this repository as a plugin marketplace, then install from it:\n\n```\n/plugin marketplace add heffrey/hordev\n/plugin install hordev@hordev\n```\n\nThe entrypoint skill loads itself at session start, so there is nothing else to configure. Restart the session after installing.\n\nRequires `bash`. Nothing else.\n\nEarly. Version 0.2.0.\n\nThe skills are written. Both hooks run correctly when executed directly —\n`session-start.sh` emits the entrypoint as valid JSON, and `reflect.sh` blocks\nonce when the run log grows, honors the `stop_hook_active` loop guard, and\nstays silent afterward. They have not yet been exercised inside a live Claude\nCode session, which is a different thing and the next thing to verify.\n\nThe library has not been run in anger across enough projects for its own self-improvement loop to have taught it anything. Expect rough edges, and expect these rules to change as real runs falsify them.\n\nMIT. See [LICENSE](/heffrey/hordev/blob/main/LICENSE).\n\nDerived in part from [superpowers](https://github.com/obra/superpowers) by\nJesse Vincent, used under the MIT License.", "url": "https://wpnews.pro/news/show-hn-hordev-claude-code-skills-that-build-instead-of-asking-questions", "canonical_source": "https://github.com/heffrey/hordev", "published_at": "2026-09-09 17:45:35+00:00", "updated_at": "2026-09-09 17:59:06.326065+00:00", "lang": "en", "topics": ["ai-tools", "ai-agents", "developer-tools"], "entities": ["Hordev", "Claude Code", "superpowers"], "alternates": {"html": "https://wpnews.pro/news/show-hn-hordev-claude-code-skills-that-build-instead-of-asking-questions", "markdown": "https://wpnews.pro/news/show-hn-hordev-claude-code-skills-that-build-instead-of-asking-questions.md", "text": "https://wpnews.pro/news/show-hn-hordev-claude-code-skills-that-build-instead-of-asking-questions.txt", "jsonld": "https://wpnews.pro/news/show-hn-hordev-claude-code-skills-that-build-instead-of-asking-questions.jsonld"}}